Munich's Helsing closed a $1.8 billion Series E on July 13 at an $18 billion valuation, the largest financing round in European defence-tech history. Combined with Quantum Systems' $1.2 billion raise eleven days earlier, Europe has now backed two battlefield-AI companies with more capital in a single month than most of its civilian AI sector raises in a year.
That capital concentration sits next to a governance moment. Scientists are pushing Brussels not to blink when GPAI enforcement powers activate August 2, while the Commission tries to build cyber-defence capability of its own. (Tech.eu)
Research & Labs
A German consortium ships the best fully open model in two languages
Soofi S, a 31.6-billion-parameter model that activates only 3.2 billion per token, now tops English and German benchmarks among fully open models, ahead of OLMo 3 32B and Apertus 70B. It was built by Fraunhofer, DFKI, TU Darmstadt and two AI companies under the KI Bundesverband, funded by Germany's economy ministry and trained on Deutsche Telekom's Munich infrastructure.
About 99% of its training data can be independently reconstructed, a transparency bar well above what any frontier lab publishes. Licensed newspaper data makes up the remaining 1.3%, which keeps it just short of the strictest European open-source standards. (The Decoder)
Mistral moves from language models into robot navigation
Mistral shipped Robostral Navigate on July 8, an 8-billion-parameter model that lets robots navigate unfamiliar environments using a single camera, no depth sensor or LIDAR required. It scored 76.6% on the R2R-CE navigation benchmark, ahead of systems using multiple sensors, and was trained entirely in simulation.
This is Mistral's fourth product ship in five weeks, after OCR 4, the sovereignty pitch aimed at former Anthropic customers, and the still-unreleased open-weight model teased for this summer. The pattern reads as a lab betting on breadth while it waits on the flagship release. (Mistral AI)
Deployment & Applications
Madrid puts AI diagnosis tools into public hospitals
Madrid's regional health service signed a €12.8 million, four-year agreement with Spain's digital ministry on July 9 to deploy AI across primary care and hospitals. The plan covers remote monitoring for chronic disease, risk-pattern analytics, and diagnostic assistance starting with early detection in primary care, building on an existing pilot that reads diabetic retinopathy scans.
The Regional Health Directorate sets rollout priorities before year-end. The money is committed; the sequencing isn't yet. (Moncloa)
A Belgian startup's AI is already cutting food waste on production lines
Polysense raised $10.7 million on July 8 for AI that inspects food production lines in real time and automatically adjusts machinery before defective batches happen, rather than flagging them after. The Ghent-area company has moved past pilots to commercial deployments across vegetable, bakery and confectionery lines in Europe, the US and the Middle East.
That's a rarer thing in industrial AI than another funding round: a system customers are actually running in production. (Tech.eu)
Finland's NestAI puts battlefield AI models into active pilots
NestAI, the defence-AI lab Silo AI founder Peter Sarlin started after selling his last company to AMD in 2024, has launched its first production models: autonomy software for edge drones and a battlefield-orchestration system called NestOS that lets militaries plan and run missions. Both are already piloting with Estonian and Finnish armed forces.
Sarlin's pitch is deliberately narrow, domain-specific defence AI rather than a general-purpose model, addressing a concern European governments have voiced for years about depending on foreign-built military software. (Sifted)
Governance & Safety
Brussels tries to get ahead of AI-powered cyberattacks
The European Commission presented an EU Action Plan on Cybersecurity and AI on July 7, acknowledging that advanced models can now find and exploit software vulnerabilities in minutes at a fraction of the cost of a human researcher. The plan commits to EU-level evaluation capacity for advanced models, a secure testing platform for critical sectors run with ENISA, and an AI-for-cybersecurity "Grand Challenge" to pool European research talent.
Rare, for an AI Act-adjacent document, to read as building capability rather than just writing rules. (European Commission)
Scientists ask Brussels not to blink on August 2
A coalition including Yoshua Bengio and Stuart Russell wrote to Commission President von der Leyen and digital chief Henna Virkkunen this week, urging the AI Office to use its new enforcement powers over general-purpose AI models in full once they activate August 2, the deadline flagged as this briefing's "one to watch" last week. The letter cites cyber-offence capabilities in current frontier models and approaching thresholds on biological risk and loss of control as reasons the moment isn't theoretical.
The AI safety research community is signalling that the AI Office's first enforcement decisions, not the legislative fight that produced them, are the real test of the Act. (The EU AI Act Newsletter)
Compute & Infrastructure
Anthropic recruits a European fintech founder to run its compute strategy
Anthropic has hired Tom Blomfield, the Monzo cofounder and former Y Combinator group partner, to work alongside Tom Brown on its compute team. Blomfield argues compute access, not model architecture, is now the binding constraint as AI systems move toward more autonomous capability.
The hire lands three weeks after Anthropic's own 19-day access blackout for European customers, a reminder that the compute Blomfield will be allocating sits mostly in US-controlled infrastructure. That's the exact dependency Mistral and Germany's SPRIND programme are trying to build an alternative to. (Sifted)
AI & Society
Allianz cuts up to 1,800 jobs as its Claude-based platform scales
Allianz Partners, the insurer's assistance and travel unit, will eliminate 1,500 to 1,800 of its roughly 22,000 European jobs, CEO Tomas Kunzmann said July 8, as AI takes on more claims and customer-service work built on the Anthropic partnership Allianz announced in January. Kunzmann has committed to severance and early-retirement packages rather than straight layoffs, calling it treating affected staff "fairly."
It's one of the clearer European test cases of what a company says "responsible" AI-driven workforce reduction looks like, made in public rather than through a quiet trickle of layoffs. (PYMNTS)
One to watch
The high-risk classification guidelines close for comment July 23
The Commission's draft guidelines on which AI systems count as "high-risk" under Article 6, covering biometrics, employment, education, credit scoring and law enforcement, take feedback until July 23, with final guidelines due by year-end. The draft already leans toward an expansive reading: more embedded and product-integrated AI systems in scope than industry expected.
Whichever way the final text lands will decide which European AI products need full conformity assessment starting December 2027 and which don't, the single biggest determinant of compliance cost still open for argument. (European Commission)